Keywords:

machine learning, graphical models, probabilistic reasoning, nonlinear models, variational methods, state-space models, hidden Markov models, inductive logic programming, first-order logic

ABSTRACT

Statistical data analysis is becoming more and more important when growing amounts of data are collected in various fields of life. Automated learning algorithms provide a way to discover relevant concepts and representations that can be further used in analysis and decision making.

Graphical models are an important subclass of statistical machine learning that have clear semantics and a sound theoretical foundation. A graphical model is a graph whose nodes represent random variables and edges define the dependency structure between them. Bayesian inference solves the probability distribution over unknown variables given the data. Graphical models are modular, that is, complex systems can be built by combining simple parts. Applying graphical models within the limits used in the 1980s is straightforward, but relaxing the strict assumptions is a challenging and an active field of research.

This thesis introduces, studies, and improves extensions of graphical models that can be roughly divided into two categories. The first category involves nonlinear models inspired by neural networks. Variational Bayesian learning is used to counter overfitting and computational complexity. A framework where efficient update rules are derived automatically for a model structure given by the user, is introduced. Compared to similar existing systems, it provides new functionality such as nonlinearities and variance modelling. Variational Bayesian methods are applied to reconstructing corrupted data and to controlling a dynamic system. A new algorithm is developed for efficient and reliable inference in nonlinear state-space models.

The second category involves relational models. This means that observations may have distinctive internal structure and they may be linked to each other. A novel method called logical hidden Markov model is introduced for analysing sequences of logical atoms, and applied to classifying protein secondary structures. Algorithms for inference, parameter estimation, and structural learning are given. Also, the first graphical model for analysing nonlinear dependencies in relational data, is introduced in the thesis.

Raiko, T. (2006):
Bayesiläinen päättely epälineaarisissa ja rakenteisissa piilomuuttujamalleissa. Tohtorin väitöskirja, Teknillinen korkeakoulu, Dissertations in Computer and Information Science, raportti D18, Espoo, Suomi.